stop() built-in/op/kernel

[pdamme]

DaphneDSL should offer a stop() built-in function which stops the execution of the script (similar to the one in Apache SystemDS). This would be very helpful, e.g., since it would allow UDFs to react to invalid arguments. In such cases, the program should not simply continue (and printing a warning could easily be overlooked), the program should stop.

Example:

def foo(x:si64) {
    if(x < 0)
        stop("x must not be less than zero");
    return 1 + sqrt(x);
}

We need:

• A stop() built-in function in DaphneDSL, which should take an optional message string that is displayed along with an error message. If no argument is provided at DSL level, an empty string could be used under the covers.
• A StopOp DaphneIR operation, with mandatory message operand.
• A stop kernel; the most straightforward way to ensure that an error message is printed and the program execution terminates in a controlled way would be to simply throw a std::runtime_error in the kernel. This exception would be caught by DAPHNE's (WIP) error handling infrastructure as any other error.
• Ideally support for stop in DaphneLib (Python API), maybe as a function on the DaphneContext.
• Unit test cases checking for an exception (CHECK_THROWS in catch2) and script-level test cases checking for unsuccessful program termination.

Hints:

• Please search for #658 in the code base (especially in the scripts/ directory), since multiple places where stop() should be used instead of a print() are marked that way.
• As a possible extension (and that could be done later), the DAPHNE compiler could detect occurrences of stop() that happen unconditionally (e.g., because the check surrounding it has been resolved at compile-time through constant propagation). In such cases, we do not even need to start the script execution, but could immediately abort with the given error message at compile-time.

[Garic152]

I'd like to work on this issue.

[pdamme]

Great, thanks. Go ahead!

[Garic152]

I've just pushed the first version of the stop() function to my fork: Fork: dev-issue-658.

The functionality is as follows:

Custom Input

File:

stop("Custom error");

Output:

[error]: Execution error: The kernel-function _stop__char failed during runtime with the following message [ System stopped: Custom Error ]
    | Source file -> "/daphne/test.daph":1:0
    |
  1 | stop("Custom Error");
    | ^~~

No Input

File:

stop();

[error]: Execution error: The kernel-function _stop__char failed during runtime with the following message [ System stopped: Unspecified error occurred. ]
    | Source file -> "/daphne/test.daph":1:0
    |
  1 | stop();
    | ^~~

When throwing the std::runtime_error, the function does show the failing lines inside the .daph file, but also mentions _stop__char. Should only the second part of the message be printed, and if so, do you have any suggestions on how to implement this?

Apart from that, if there aren't any major concerns in how I've implemented this functionality, I'll continue working on the Python API implementation as well as the unit/script tests.

[pdamme]

Thanks for the update. The error message looks fine as it is for now. The overall format is generated by our error handling facilities. That includes the kernel name (_stop__char), because it's useful information for users and developers. Furthermore, that way, the error message is consistent with other errors. So having the message from the stop() function inside the general error message is totally fine.

Regarding the Python API: Thinking about it again now, adding the stop() function there will most likely not work out at the moment. This is because stop() doesn't have any results (values it returns). Therefore, it will not be included in the generated DaphneDSL script. That is a general limitation of DaphneLib at the moment (see #712), which we will fix later. Thus, you can skip the DaphneLib part for this issue.

The code in your fork looks very good already. The main thing that's still missing are indeed the test cases.

[Garic152]

The stop() function now works like requested and also includes the unit/script-level tests. The script-level tests are located in ControlFlowTest.cpp, as this seemed like the most logical place.

When running ./test.sh -nb -d yes, I encountered errors in DistributedTest and MatrixLiteralTest. These errors should be unrelated to my recent commit because I didn't change any files related to these tests. If there is a correlation though, please let know.

If there are any issues with my implementation, please let me know.

[pdamme]

Thanks for the update. At a quick glance, the tests in your fork look good. It would be great if you could employ the stop() built-in function in the DaphneDSL scripts as mentioned in the hints above. Other than that, feel free to create a pull request :) .

Regarding the errors: the DistributedTest also sometimes fails on my system (for unknown reasons), but regarding the MatrixLiteralTest, I'm not aware of any problems. But thanks for letting us know; sometimes there are non-deterministic errors that do not occur in all environments. I agree that your changes are unrelated, though.